ClawHub
Back to skill

Security audit

huawei-cloud-cloudrobo-train

Security checks for vulnerabilities and agentic risk

Overview

The skill is coherent for managing Huawei Cloud CloudRobo training jobs, but it instructs agents to launch or resubmit costly long-running work without a final confirmation prompt.

Review before installing. Use this skill only with least-privilege CloudRobo credentials and an explicit workspace ID, and require a human confirmation step before create-task, restart-task, resume-task, stop-task, delete-tasks, save-draft, register-checkpoint, or any command that can spend compute resources or alter tasks. Be careful running the included test script because it can create drafts, restart them, and delete tasks after only a brief terminal warning.

A.I.G

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs

SkillSpector

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (47)

Ae1

High
Category
analysis-evasion
Content
> `references/task-config-catalog.md` for the coverage matrix.
Confidence
100% confidence
Finding
Referenced artifact was not completely inspected

Ae1

High
Category
analysis-evasion
Content
> `references/task-config-catalog.md` for the coverage matrix.
Confidence
100% confidence
Finding
Referenced artifact was not completely inspected

Tool Parameter Abuse

High
Category
Tool Misuse
Content
| List tasks | `list-tasks [--train-mode] [--status] [--offset] [--limit]` | `list_train_tasks` / `list_sim_rl_tasks` | `GET /` |
| Show task | `show-task --task-id <id>` | `show_train_task` / `show_sim_rl_task` | `GET /{task_id}` |
| Update task | `update-task --task-id <id> --config '<json>'` | `update_train_task` / `update_sim_rl_task` | `PATCH /{task_id}` |
| Delete tasks | `delete-tasks --task-id <id> [--task-id <id>...]` | `batch_delete_train_tasks(execution_ids)` / `delete_sim_rl_task(task_id)` | `POST /batch-delete` (train) / `DELETE /{task_id}` (SimRL) |

> **delete-tasks behavior**: For regular training tasks, the CLI auto-resolves `execution_id` from
> the provided `task_id` via `show-task` before calling batch-delete. Users can pass task IDs
Confidence
80% confidence
Finding
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Ae1

High
Category
analysis-evasion
Content
- [API Paths](references/api-paths.md) — Full endpoint list with SDK source line references
Confidence
100% confidence
Finding
Referenced artifact was not completely inspected

Ae1

High
Category
analysis-evasion
Content
- **Test**: `bash scripts/test-cli-commands.sh` (CLI/SDK/API); see `templates/test-vars.json` for full coverage
Confidence
100% confidence
Finding
Referenced artifact was not completely inspected

Tool Parameter Abuse

High
Category
Tool Misuse
Content
| Save draft (train / SimRL) | Write access to `cloudrobo-service` (`/v1/training/train-tasks/draft`, `/v1/training/rl-tasks/simulation/draft`) |
| Update task name/description | Write access to `cloudrobo-service` (`PATCH /v1/training/train-tasks/{id}`, `PATCH /v1/training/rl-tasks/simulation/{id}`) |
| Batch delete train-tasks | Delete access to `cloudrobo-service` (`POST /v1/training/train-tasks/batch-delete`) |
| Delete SimRL task | Delete access to `cloudrobo-service` (`DELETE /v1/training/rl-tasks/simulation/{id}`) |
| View stages / resource-usage / events / logs / signed-url | Read access to task sub-resources |
| Count tasks by status | Read access to `cloudrobo-service` (`/v1/training/train-tasks/stats`, `/v1/training/rl-tasks/simulation/stats`) |
| Checkpoint management (list/register) | Read/Write access to `cloudrobo-service` (`/v1/training/train-tasks/{id}/checkpoints`) |
Confidence
80% confidence
Finding
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Tool Parameter Abuse

High
Category
Tool Misuse
Content
| create_sim_rl_task_draft | `create_sim_rl_task_draft(req)` | `save-draft --sim-rl` | `POST /v1/training/rl-tasks/simulation/draft` |
| show_sim_rl_task | `show_sim_rl_task(task_id)` | `show-task --sim-rl` | `GET /v1/training/rl-tasks/simulation/{task_id}` |
| update_sim_rl_task | `update_sim_rl_task(task_id, req)` | `update-task --sim-rl` | `PATCH /v1/training/rl-tasks/simulation/{task_id}` |
| delete_sim_rl_task | `delete_sim_rl_task(task_id)` | `delete-tasks --sim-rl` | `DELETE /v1/training/rl-tasks/simulation/{task_id}` |
| stop_sim_rl_task | `stop_sim_rl_task(task_id)` | `stop-task --sim-rl` | `POST /v1/training/rl-tasks/simulation/{task_id}/stop` |
| copy_sim_rl_task | `copy_sim_rl_task(task_id, req=None, task_detail=None)` | `clone-task` | `POST /v1/training/rl-tasks/simulation/{task_id}/copy` |
| restart_sim_rl_task | `restart_sim_rl_task(task_id, req=None, workspace_id=None, task_detail=None)` | `restart-task --sim-rl` | `POST /v1/training/rl-tasks/simulation/{task_id}/restart` |
Confidence
80% confidence
Finding
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Vague Triggers

Medium
Confidence
94% confidence
Finding
The trigger list is extremely broad and includes generic concepts like model training, logs, events, and statistics in both English and Chinese. That increases the chance the skill is invoked unintentionally in ordinary conversation, which is risky because this skill can create, restart, stop, and delete cloud training jobs with cost and operational impact.

Missing User Warnings

Medium
Confidence
97% confidence
Finding
The skill explicitly instructs silent submission of task creation and restart flows even though the document repeatedly notes these jobs can run for hours or days and consume cloud resources. A mistaken or induced invocation could therefore launch costly, long-running training workloads without a clear user acknowledgment at the point of action.

Autonomous Decision Making

Medium
Category
Excessive Agency
Content
and `algorithm_version_id` (2-field format). The backend auto-resolves engine/command from the
     algorithm asset. Top-level `inputs`/`outputs` arrays are required. See [Workspace Model Sub-flow](#step-3a-ws--workspace-model-sub-flow) below.
2. **Select model** → extract `asset_id` + `latest_version_id` from the selected model's list entry.
   **Do NOT ask user for version again** — auto-use `latest_version_id`. Only ask if user explicitly
   wants a non-latest version.
3. **Query model version detail** `GET /v1/assets/{asset_id}/versions/{version_id}` → get `actions`
   array. Each action: `{action, algorithm:{asset_id, version_id}, status}`
Confidence
80% confidence
Finding
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Autonomous Decision Making

Medium
Category
Excessive Agency
Content
and `algorithm_version_id` (2-field format). The backend auto-resolves engine/command from the
     algorithm asset. Top-level `inputs`/`outputs` arrays are required. See [Workspace Model Sub-flow](#step-3a-ws--workspace-model-sub-flow) below.
2. **Select model** → extract `asset_id` + `latest_version_id` from the selected model's list entry.
   **Do NOT ask user for version again** — auto-use `latest_version_id`. Only ask if user explicitly
   wants a non-latest version.
3. **Query model version detail** `GET /v1/assets/{asset_id}/versions/{version_id}` → get `actions`
   array. Each action: `{action, algorithm:{asset_id, version_id}, status}`
Confidence
80% confidence
Finding
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Autonomous Decision Making

Medium
Category
Excessive Agency
Content
and `algorithm_version_id` (2-field format). The backend auto-resolves engine/command from the
     algorithm asset. Top-level `inputs`/`outputs` arrays are required. See [Workspace Model Sub-flow](#step-3a-ws--workspace-model-sub-flow) below.
2. **Select model** → extract `asset_id` + `latest_version_id` from the selected model's list entry.
   **Do NOT ask user for version again** — auto-use `latest_version_id`. Only ask if user explicitly
   wants a non-latest version.
3. **Query model version detail** `GET /v1/assets/{asset_id}/versions/{version_id}` → get `actions`
   array. Each action: `{action, algorithm:{asset_id, version_id}, status}`
Confidence
80% confidence
Finding
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Autonomous Decision Making

Medium
Category
Excessive Agency
Content
and `algorithm_version_id` (2-field format). The backend auto-resolves engine/command from the
     algorithm asset. Top-level `inputs`/`outputs` arrays are required. See [Workspace Model Sub-flow](#step-3a-ws--workspace-model-sub-flow) below.
2. **Select model** → extract `asset_id` + `latest_version_id` from the selected model's list entry.
   **Do NOT ask user for version again** — auto-use `latest_version_id`. Only ask if user explicitly
   wants a non-latest version.
3. **Query model version detail** `GET /v1/assets/{asset_id}/versions/{version_id}` → get `actions`
   array. Each action: `{action, algorithm:{asset_id, version_id}, status}`
Confidence
80% confidence
Finding
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Whitespace Padding

Medium
Category
Prompt Injection
Content
**Algorithm config differences:**

| Sub-path | Algorithm format                                                                                                                               | Discovery |
|----------|------------------------------------------------------------------------------------------------------------------------------------------------|-----------|
| Gallery (预制) | 2 fields: `algorithm_asset_id` + `algorithm_version_id`. Backend auto-resolves engine/image.                                                   | `list-publication-assets --type algorithm` → extract `algorithm_asset_id` + `latest_version_id`. Query `ext_metadata` for hyperparams/env/resource. |
| Workspace (空间资产) | 7 fields: `engine.image_url` + `image_asset_id` + `image_version_id` + `code_dir` + `command` + `algorithm_asset_id` + `algorithm_version_id`. | `list-assets --type algorithm` → extract `algorithm_asset_id` + `latest_version_id`. Query `ext_metadata` for engine/command/code_dir + hyperparams/env/resource. |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
## Edge Cases

| Scenario | Handling                                                                                                                                                                                                                                                 |
| ---------- |----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Missing `workspace_id` | All commands auto-resolve from config or auto-query; run `cloudrobo workspace use` to set default                                                                                                                                                        |
| Task in non-terminal state | Poll at 30-60s intervals; training can run for hours/days                                                                                                                                                                                                |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| Scenario | Handling                                                                                                                                                                                                                                                 |
| ---------- |----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Missing `workspace_id` | All commands auto-resolve from config or auto-query; run `cloudrobo workspace use` to set default                                                                                                                                                        |
| Task in non-terminal state | Poll at 30-60s intervals; training can run for hours/days                                                                                                                                                                                                |
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| Scenario | Handling                                                                                                                                                                                                                                                 |
| ---------- |----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Missing `workspace_id` | All commands auto-resolve from config or auto-query; run `cloudrobo workspace use` to set default                                                                                                                                                        |
| Task in non-terminal state | Poll at 30-60s intervals; training can run for hours/days                                                                                                                                                                                                |
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| Scenario | Handling                                                                                                                                                                                                                                                 |
| ---------- |----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Missing `workspace_id` | All commands auto-resolve from config or auto-query; run `cloudrobo workspace use` to set default                                                                                                                                                        |
| Task in non-terminal state | Poll at 30-60s intervals; training can run for hours/days                                                                                                                                                                                                |
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
| CREATE_FAILED | Check `spec` format, `cluster_id`, resource availability, task name uniqueness (409 Conflict), **input model status** (must be `DRAFT` for workspace models)                                                                                             |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| ---------- |----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Missing `workspace_id` | All commands auto-resolve from config or auto-query; run `cloudrobo workspace use` to set default                                                                                                                                                        |
| Task in non-terminal state | Poll at 30-60s intervals; training can run for hours/days                                                                                                                                                                                                |
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
| CREATE_FAILED | Check `spec` format, `cluster_id`, resource availability, task name uniqueness (409 Conflict), **input model status** (must be `DRAFT` for workspace models)                                                                                             |
| RUN_FAILED | Check logs/events; common: OOM, image pull failure, dataset access denied, **dataset format incompatibility** (exitCode 1 in <2 min, no logs)                                                                                                            |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| Missing `workspace_id` | All commands auto-resolve from config or auto-query; run `cloudrobo workspace use` to set default                                                                                                                                                        |
| Task in non-terminal state | Poll at 30-60s intervals; training can run for hours/days                                                                                                                                                                                                |
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
| CREATE_FAILED | Check `spec` format, `cluster_id`, resource availability, task name uniqueness (409 Conflict), **input model status** (must be `DRAFT` for workspace models)                                                                                             |
| RUN_FAILED | Check logs/events; common: OOM, image pull failure, dataset access denied, **dataset format incompatibility** (exitCode 1 in <2 min, no logs)                                                                                                            |
| Resource scheduling failure | Check `spec`, `worker_num`, cluster capacity, and `ext_metadata.resource` min NPU constraint                                                                                                                                                             |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| Missing `workspace_id` | All commands auto-resolve from config or auto-query; run `cloudrobo workspace use` to set default                                                                                                                                                        |
| Task in non-terminal state | Poll at 30-60s intervals; training can run for hours/days                                                                                                                                                                                                |
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
| CREATE_FAILED | Check `spec` format, `cluster_id`, resource availability, task name uniqueness (409 Conflict), **input model status** (must be `DRAFT` for workspace models)                                                                                             |
| RUN_FAILED | Check logs/events; common: OOM, image pull failure, dataset access denied, **dataset format incompatibility** (exitCode 1 in <2 min, no logs)                                                                                                            |
| Resource scheduling failure | Check `spec`, `worker_num`, cluster capacity, and `ext_metadata.resource` min NPU constraint                                                                                                                                                             |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| Task in non-terminal state | Poll at 30-60s intervals; training can run for hours/days                                                                                                                                                                                                |
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
| CREATE_FAILED | Check `spec` format, `cluster_id`, resource availability, task name uniqueness (409 Conflict), **input model status** (must be `DRAFT` for workspace models)                                                                                             |
| RUN_FAILED | Check logs/events; common: OOM, image pull failure, dataset access denied, **dataset format incompatibility** (exitCode 1 in <2 min, no logs)                                                                                                            |
| Resource scheduling failure | Check `spec`, `worker_num`, cluster capacity, and `ext_metadata.resource` min NPU constraint                                                                                                                                                             |
| Stopped task | Use `restart-task` to resubmit; `resume-task` for supported train-only cases                                                                                                                                                                             |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
| CREATE_FAILED | Check `spec` format, `cluster_id`, resource availability, task name uniqueness (409 Conflict), **input model status** (must be `DRAFT` for workspace models)                                                                                             |
| RUN_FAILED | Check logs/events; common: OOM, image pull failure, dataset access denied, **dataset format incompatibility** (exitCode 1 in <2 min, no logs)                                                                                                            |
| Resource scheduling failure | Check `spec`, `worker_num`, cluster capacity, and `ext_metadata.resource` min NPU constraint                                                                                                                                                             |
| Stopped task | Use `restart-task` to resubmit; `resume-task` for supported train-only cases                                                                                                                                                                             |
| SimRL resume | Not supported; `--sim-rl` not accepted on `resume-task`                                                                                                                                                                                                  |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| `spec` format | String `Ascend: <n> * <model> \| <vCPUs> vCPUs \| <GiB> GiB`, not JSON. Filter by `ext_metadata.resource` constraints                                                                                                                                    |
| `train_method` / `train_mode` | Uppercase enums: `FFT`/`SFT`/`LORA`/`QLORA`/`DEEPSPEED`; `MODEL_TUNING`/`TRAIN_FROM_SCRATCH`                                                                                                                                                             |
| CREATE_FAILED | Check `spec` format, `cluster_id`, resource availability, task name uniqueness (409 Conflict), **input model status** (must be `DRAFT` for workspace models)                                                                                             |
| RUN_FAILED | Check logs/events; common: OOM, image pull failure, dataset access denied, **dataset format incompatibility** (exitCode 1 in <2 min, no logs)                                                                                                            |
| Resource scheduling failure | Check `spec`, `worker_num`, cluster capacity, and `ext_metadata.resource` min NPU constraint                                                                                                                                                             |
| Stopped task | Use `restart-task` to resubmit; `resume-task` for supported train-only cases                                                                                                                                                                             |
| SimRL resume | Not supported; `--sim-rl` not accepted on `resume-task`                                                                                                                                                                                                  |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Whitespace Padding

Medium
Category
Prompt Injection
Content
| CREATE_FAILED | Check `spec` format, `cluster_id`, resource availability, task name uniqueness (409 Conflict), **input model status** (must be `DRAFT` for workspace models)                                                                                             |
| RUN_FAILED | Check logs/events; common: OOM, image pull failure, dataset access denied, **dataset format incompatibility** (exitCode 1 in <2 min, no logs)                                                                                                            |
| Resource scheduling failure | Check `spec`, `worker_num`, cluster capacity, and `ext_metadata.resource` min NPU constraint                                                                                                                                                             |
| Stopped task | Use `restart-task` to resubmit; `resume-task` for supported train-only cases                                                                                                                                                                             |
| SimRL resume | Not supported; `--sim-rl` not accepted on `resume-task`                                                                                                                                                                                                  |
| Draft submit via CLI | `restart-task` supports `--config`/`--config-file` to edit fields before resubmit                                                                                                                                                                        |
| Missing required fields | SDK validates before HTTP call; CLI shows `click.UsageError` with the missing field list                                                                                                                                                                 |
Confidence
70% confidence
Finding
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Static analysis

No suspicious patterns detected.